Zuzana Burivalova, University of Wisconsin
Video Recording
Abstract:
Conservation and sustainable management projects need to be able to monitor their progress – or lack thereof – in a timely and rigorous way that is scalable: increasing biodiversity with forest regeneration in Indonesia should be possible to measure with the same indicator as in Myanmar. Soundscape monitoring can capture vocalizing diversity over time and across sites, and is being increasingly tested as a tool for conservation monitoring. A major obstacle towards this is the lack of knowledge of how natural and human modified soundscapes change across scales. In this talk, I will describe how soundscapes change between land use types, as well as with different intensities within land use types. I will introduce three soundscape analysis techniques, on examples of soundscapes from Myanmar, Indonesia, and Papua New Guinea.
Bio:
Dr. Zuzana Buřivalová is a conservation scientist with a goal to find ways in which we can protect tropical forest biodiversity in an equitable way. Zuzana is an Assistant Professor at the University of Wisconsin-Madison, where she is affiliated with the department of Forest & Wildlife Ecology, The Nelson Institute for Environmental Studies, and the Center for Sustainability and the Global Environment (SAGE). Zuzana is the Principal Investigator of the Sound Forest Lab, a group of researchers using soundscapes - all the sounds that can be heard in a landscape - to understand the health of rainforests. The Sound Forest Lab collaborates with NGOs, governments, and local communities to design research projects that can help on the ground.
Summary:
Goal: measure forest biodiversity in tropical forests
Traditional biodiversity estimation:
Vegetation map of an area
Aerial photography
Surveys
Lots of guesswork
Accuracy of this has improved significantly as quality of satellite and drone imagery has improved
Major limitation: images rarely capture animal species
Focus of talk:
Use of audio data to identify animals from the sounds they make
Mammals, birds, frogs can be captured in this way
Bioacoustic workflow
Forest soundscapes produce sound
Sensors collect audio
Recording devices left in the forest for a long time
Battery operated
Hopefully won’t break due to moisture or infestation
Data processing
Ecological analysis:
Soundscape indices: quantify biodiversity level without understanding the species
E.g. Complexity of sound (how much it changes)
More changes suggest more complexity
But unclear if changes are due to real species, rain, etc.
Individual species analysis
This is a direct measurement of diversity and species abundance
High cost per species
Some species have unknown sounds (especially, insects and frogs)
Whole soundscape analysis
Evaluate the health of forests,
Baseline
After interventions: logging, management, reforestation
E.g. a reforestation project may store carbon but does the new forest promote a healthy ecology?
The index must work in any tropical forest across the world
There are many soundscape indices (~60) but all have issues
New index: soundscape saturation
Plot time of day vs frequency of sound
Intuition:
Different species use different regions of the audio (frequency/time) space so their sound isn’t drowned out
Can identify that species are missing by gaps in the soundscape
Analysis plots the presence of noise above a threshold in each point in the time-of-day (by minute) / frequency plot (threshold controls for loud species very close to microphone or weak background noise)
Can observe distribution of sounds over the day (dawn and dusk are very populated with sounds)
Approach is calibrated by experts labeling different sounds by their likely species (sonotypes)
One species can make one, but maybe multiple sounds
Can use this to count the number of different sonotypes in the data
Soundscapes explains 30% of variation in sonotype data
Different types of forests with different known levels of biodiversity have very different soundscape saturation index values
Overall values are different
Very large difference at dawn and dusk between healthy forests and plantations (loudest, most active times)
Can compare healthy forests to different types of logging
Some types of logging have a much smaller impact on species
Taxonomic group analysis
Separation of the soundscape into birds, insects and amphibians
Birds and insects are easier to divide
Frogs and mammals are less so
Individual species analysis
Neural networks trained on sounds from known species
Neural networks need a lot of data, which limits their applicability
E.g. BirdNet can identify ~500 species of common birds in North America and Europe
Not applicable to less studied species (e.g. in Southeast Asia)
Need a model that
Requires little training data
Open source
Can classify all animal sounds, from various taxonomic groups
Reasonable accuracy (>80%)
Good enough to identify sounds as coming from different species, even if the exact species is unknown
Challenge: dataset is very imbalance
Few species have many examples
Many species have few examples (<=3)
Approach:
Data augmentation by mutating the existing dataset (stretching, adding rain sounds)
Transfer learning across species
Both techniques were critical for achieving adequate accuracy
Accuracy was good across all taxonomic groups
Works with 3 samples (good minimum: train, validate, test)
Challenges:
Hard to segment spectrogram to identify the sounds to be identified
How do we deal with sounds that don’t belong to any known classes?
Need to make the algorithm runnable by people who don’t know how to code